How it fits together

How the three parts of the platform work together when a developer builds an agent, when a user makes a request, and when an administrator changes the configuration.

Building an agent

1. Drafteditor or Draft with AI2. Validatechecks, then a traced run3. Publisha new version4. Describeindexed for searchA change to a published agent starts from a new draft
From a draft to a searchable agent
  1. A developer creates a draft in the UI and writes the agent definition in the code editor or with Draft with AI. A draft is private to the developer who created it.
  2. The developer validates the draft. The cloud service checks the definition and every Python expression in it. The agent server then runs the draft against the developer's validation query with tracing on, so that every step is visible.
  3. The developer publishes the validated draft. Every developer can see the published agent, and users can run it. A publish that changes the definition creates a new version.
  4. The developer writes a plain-English description of the agent. The cloud service indexes the description in 15 to 30 seconds, and search can then find the agent.

A change to a published agent follows the same steps, starting from a new draft.

Handling a request

1. Requesta user asks2. Matchabout 200 ms3. RunLLMs and systems4. Outputstreamed5. Savestate, encryptedA follow-up request restores the state and continues the conversation
From a request to an answer
  1. A user makes a request in the UI or in a chat application. An agent harness or your own program makes requests in the same way, through the REST API.
  2. The agent server sends the request to the cloud service. Search matches the meaning of the request to the agent descriptions in about 200 milliseconds. It returns one agent, two when the request is ambiguous, or none when no agent fits.
  3. The agent server runs the matched agent in the agent runtime. The agent calls LLMs and your systems, and can pause to ask the user for input.
  4. The agent server streams the output to the user as the agent runs.
  5. When the agent ends or pauses, the agent server encrypts the conversation state and saves it in the cloud service. An account can keep the state in its own organization instead.

A follow-up request repeats these steps in the same conversation. The state is restored, and the same agent or a different one runs with it. This is how agents build on each other's work.

Keeping agent servers current

UIchangestoreAgent server AstatelessAgent server BstatelessAgent server CstatelessCloud serviceholds all stateBefore every agent run:1. report the run to the cloud service2. learn the time of the latest change3. fetch only what changed4. rebuild the runtime, then runA change is live on the next run on every agent server. A running agent keeps the runtime it started with.
How a configuration change reaches every agent server

Configuration, such as profiles, the allowlist and the runtime limits, is edited in the UI. The UI sends each change to the agent server, and the agent server stores it in the cloud service. Each time an agent run starts, the agent server reports the run to the cloud service, and the reply carries the time of the latest change. When the configuration has changed, the agent server fetches only what changed and rebuilds its runtime before the run.

A change is live on the next run on every agent server, and two agent servers that start the same agent at the same moment run with the same configuration. A running agent keeps the runtime it started with. See Agent runtime.